Research 02 · Under review
QIH-RAG
An honest, controlled study of which parts of a biomedical retrieval pipeline actually matter.
The problem
Language models in clinical workflows make up plausible facts, which is a safety problem rather than a benchmark nuisance. Retrieval-augmented generation helps, but biomedical text breaks the usual recipe.
Generic chunkers split dosage instructions and contraindication lists. Dense retrieval misses exact terminology while BM25 misses paraphrases. And the configuration space is too large to tune by hand.
What we built
- 01MedSABC chunker
Keeps eight types of clinical spans, such as dosages and criteria, intact.
- 02Hybrid retrieval
BM25 and dense BGE scores fused with query-type-adaptive weights.
- 03Cross-encoder rerank
A second stage re-scores the shortlist for precision.
- 04Automated config search
Quantum-inspired search, PSO and GA under an identical 72-evaluation budget.
Results
All three systems retrieve at the same chunk granularity, so this is a like-for-like comparison.
What we found
Hybrid retrieval is the lever that matters
Removing it costs 9.0% nDCG@10 (p = 0.020), and the effect reproduces on all three datasets. It is the only statistically significant component.
Automation helps; the quantum-inspired heuristic does not
All three optimizers beat hand-tuned BM25, but they are statistically indistinguishable from each other (p > 0.25). The paper reports this null result plainly.
It transfers
The PSO configuration reaches 0.751 nDCG@10 on SciFact zero-shot, above published SPLADE-v3 (0.710) and ColBERT v2 (0.716).
General beats specialised embeddings here
General-purpose BGE embeddings outperform the domain-specific MedCPT in the hybrid setting. The clinical chunker showed no significant aggregate effect, so it is presented as a safeguard, not a gain.
The value came from automating the search, not from the clever optimizer, and the paper reports that plainly.
A medical image search system that knows when to say “I don’t know” and hands the case to a clinician.